Neural Network Training via Motion-Based Pixel Filtering
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current scene filtering methods in autonomous vehicles using motion estimation face inefficiencies due to training neural networks on irrelevant camera data, leading to longer training times and reduced performance.
Innovation Solution
Implementing a method that identifies and filters pixels based on motion relative to the vehicle from camera data, thereby excluding stationary or unimportant data points, allowing the neural network to focus on relevant information for autonomous vehicle control operations.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Quantity of substance
If neural networks are trained on all camera data including stationary pixels, then the training dataset is larger and more comprehensive, but training time increases and performance decreases
Solution Approach 1:
The patent extracts and removes stationary pixels from the camera data before training the neural network. Motion estimation techniques identify pixels that do not move between frames, and these stationary pixels are excluded from the training dataset. This extraction of irrelevant data reduces the training dataset size while maintaining the quality of relevant moving objects, thereby decreasing training time without sacrificing performance.
2Quantity of substance
If neural networks are trained on all camera data including stationary pixels, then the training dataset is more comprehensive, but training performance and accuracy decrease
Solution Approach 1:
The patent applies motion estimation to extract and remove stationary pixels from the camera data. By identifying pixels that remain unchanged between consecutive frames, the system separates relevant moving objects from irrelevant stationary background elements. This extraction process maintains the completeness of the training dataset for moving objects while eliminating noise from stationary pixels, thereby improving training performance and accuracy.
3Productivity
If motion estimation filtering is applied to camera data, then training efficiency improves and training time decreases, but the device complexity increases
Solution Approach 1:
The patent applies motion estimation as a preliminary filtering step before neural network training. By performing motion estimation upfront to identify and remove stationary pixels from the camera data, the system prepares a pre-filtered dataset that is optimized for training. This preliminary action reduces the computational burden during the actual training phase, improving training efficiency while managing overall system complexity through staged processing.
Data Source
AI summary
Selective training of neural networks using motion estimation, including: selecting, from a plurality of pixels in video data from a vehicle, based on motion relative to the vehicle, one or more pixels; and training a neural network based on the video data by zeroing out an error function applied to the selected one or more pixels.